Translating Lexical Semantic Relations: The First Step Towards Multilingual Wordnets* Chu-Ren Huang, I-Ju E. Tseng, Dylan B.S. Tsai Institute of Linguistics, Preparatory Office, Academia Sinica 128 Sec.2 Academy Rd., Nangkang, Taipei, 115, Taiwan, R.O.C. [email protected], {elanna, dylan}@hp.iis.sinica.edu.tw Abstract Wordnets express ontology via a network of Establishing correspondences between words linked by lexical semantic relations. Since wordnets of different languages is essential these words are by definition the lexicon of each to both multilingual knowledge processing language, the wordnet design feature ensures and for bootstrapping wordnets of versatility in faithfully and comprehensively low-density languages. We claim that such representing the semantic content of each correspondences must be based on lexical semantic relations, rather than top ontology language. Hence, on one hand, these conceptual or word translations. In particular, we define atoms reflect linguistic idiosyncrasies; on the a translation equivalence relation as a other hand, the lexical semantic relations (LSR’s) bilingual lexical semantic relation. Such receive universal interpretation across different relations can then be part of a logical languages. For example, the definition of entailment predicting whether source relations such as synonymy or hypernymy is language semantic relations will hold in a universal. The universality of the LSR’s is the target language or not. Our claim is tested foundation that allows wordnet to serve as a with a study of 210 Chinese lexical lemmas potential common semantic network and their possible semantic relations links representation for all languages. The premise is bootstrapped from the Princeton WordNet. The results show that lexical semantic tacit in Princeton WordNet (WN), EuroWordNet relation translations are indeed highly precise (EWN, Vossen 1998), and MultiWordNet (MWN, when they are logically inferable. Pianta et al. 2002). It is also spelled out explicitly in the adaptation of LSR tests for Chinese 1. Introduction (Huang et al. 2001). A semantic network is critical to knowledge Given that LSR’s are semantic primitives processing, including all NLP and Semantic Web applicable to all language wordnets, and that the applications. The construction of semantic solution to the low-density problem in building networks, however, is notoriously difficult for language wordnets must involve bootstrapping ‘small’ (or ‘low-density’) languages. For these from another language, LSR’s seem to be the languages, the poverty of language resources, natural units for such bootstrapping operations. and the lack of prospect of material gains for The rich and structured semantic information infrastructure work conspire to create a vicious described in WN and EWN can be transported circle. This means that the construction of a through accurate translation if the conceptual semantic network for any small language must relations defined by LSRs remain constant in start from scratch and faces inhibitive financial both languages. In practice, such an application and linguistic challenges. would also serve the dual purpose of creating a In addition, semantic networks serve as bilingual wordnet in the process. reliable ontolog(ies) for knowledge processing In this paper, we will examine the validity only if their conceptual bases are valid and of cross-lingual LSR inferences by bootstrapping logically inferable across different languages. a Chinese Wordnet with WN. In practice, this Take wordnets (Fellbaum 1998), the de facto small-scale experiment shows how a wordnet for standard for linguistic ontology, for example. a low-density language can be built through * An earlier version of this paper was presented at the Third Chinese Lexical Semantics Workshop at Academia Sinica in May 2002. We are indebted to the participants as well as colleagues at CKIP for their comments. We would also like to thank the SemaNet 2002 reviewers for their helpful comments. It is our own responsibilities that, due to the short revision time, we were not able to incorporate all their suggestions, especially comparative studies with some relative GWA papers. We are also responsible for all remaining errors bootstrapping from an available wordnet. In the relation y, between CW1 and CW2 is a theoretical terms, we explore the logical functional combination of the three LSR’s i, x, conditions for the cross-lingual inference of and ii. LSR's. However, it is well known that language translation involves more than semantic 2. Translation Equivalents and Semantic correspondences. Social and cultural factors also Relations play a role in (human) choices of translation Note that two translation equivalents (TE) equivalents. It is not the aim of this paper to in a pair of languages stand in a lexical semantic predict when or how these semantically relation. The most desirable scenario is that when non-identical translations arise. The aim is to see the two TE’s are synonymous, such as the how much lexical semantic information is English ‘apple’ to the Mandarin ‘ping2guo3’. inferable across different languages, regardless of However, since the conceptual space is not translational idiosyncrasies. In this model, the segmented identically for all languages, TE’s prediction relies crucially on the semantic may often stand in other relations to each other. information provided by the source language (e.g. For instance, the Mandarin ‘zuo1zhi5’ is a English) lexical entry as well as the lexical hypernym for both the English ‘desk’ and ‘table’. semantic correspondence of a target language Suppose we postulate that the LSR’s between (e.g. Chinese) entry. The translation relations of TE’s are exactly identical in nature to the the relational target pairs, although capable of monolingual LSR’s described in wordnets. This introducing more idiosyncrasies, are not directly means that the lexical semantic relation involved in the prediction. Hence we make the introduced by translation can be combined with generalization that any discrepancy introduced at monolingual LRS’s. Predicting LSR’s in a target this level does not affect the logical relation of language based on source language data become LSR prediction and adopt a working model a simple logical operation of combining described in Diagram 2. We only take into relational functions when the LSR of translation consideration those cases where the translation equivalency is defined. This framework is LSR ii is exactly equivalent, i.e., EW2 = CW2. illustrated in Diagram 1. This step also allows us to reduce the maximal number of LSR combination in each prediction ¢¤£ to two. Thus we are able to better predict the CW ii ¡ 2 2 contribution of each mono- or bi-lingual LSR. ¢¤£ y x ¡ 2 = CW2 (ii = 0) ¢¤£ ¡ CW1 ¥ i 1 y x x = EW1 - EW2 LSR ¢¤£ ¡ ¥ i 1 y = CW1- CW2 LSR i = CW1 - EW1 Translation LSR The unknown LSR y = i + x ii = CW2 - EW2 Translation LSR Diagram 2. Translation-mediated LSR Prediction The unknown LSR y = i + x + ii (Reduced Model, currently adopted) Diagram 1. Translation-mediated LSR Prediction 2.1 LRS Inference as Relational Combination (The complete model) With the semantic contribution of the translation equivalency defined as a (bilingual) CW1 represents our starting Chinese lemma LSR, the inference of LSR in the target language which can be linked to EW1 through the wordnet is a simple combination of semantic translation LSR i. The linked EW1 can than relations. The default and ideal situation is where provide a set of LSR predictions based on the the two TE’s are synonymous. English WN. Assume that we take the LSR x, which is linked to EW2. That LSR prediction is mapped back to Chinese when EW2 is translated to CW2 with a translation LSR ii. In this model, ¢¡¤£ 2 = EW2 English synsets. The occurring distribution is as follows: 84 N’s with 195 times; 41 V’s with 161 times; 10 Adj’s with 47 times; and 47 Adv’s with y x 94 times. 441 distinct English synsets are covered under this process, since some of the TE’s are for the same synset. This means that ¥ i ¦ each input Chinese lemma linked to 2.4 English CW1 = EW1 (i = 0) synsets in average. Based on the TEDB and The unknown LSR y = x English WN, the 179 mapped input Chinese Diagram 3. Translation-mediated LSR Prediction lemmas expanded to 597 synonyms. And (when TE’s are synonymous) extending from the 441 English synsets, there are 1056 semantically related synsets in WN, which In this case, the translation LSR is an identical yields 1743 Chinese words with our TEDB. relation; the LSR of the source language wordnet can be directly inherited. This is illustrated in 3.1. Evaluation of the Semantics of Translation Diagram 3. Six evaluative tags are assigned for the However, when the translation has a TEDB. Four of them are remarks for future non-identical semantic relation, such as processing. The LSR marked are antonyms and hypernyms, then the LSR § predicted is the combination of the bilingual Synonymous: TE’s that are semantically relation and the monolingual relation. In this equivalent. paper, we will concentrate on Hypernyms and ¨ Other Relation: TE’s that hold other Hyponyms. The choice is made because these semantic relations two LSR’s are transitive relations by definition and allows clear logical predications when The result of evaluation of TE’s involving the combined. The same, with some qualifications, 210 chosen lemma are given in Table 1. may apply to the Holonym relations. Combinations of other LSR’s may not yield clear logical entailments. The scenarios involving Other Syn. Incorrect Total Hyponymy and Hypernymy will be discussed in Relation section 3.3. 148 32 15 195 N 3. Cross-lingual LSR Inference: A Study 75.90% 16.41% 7.69% 100% based on English-Chinese Correspondences 113 29 19 161 V In this study, we start with a WN-based 70.18% 18.01% 11.8% 100% English-Chinese Translation Equivalents 1 39 8 0 47 Database (TEDB) .